Anthropic · Proprietary · tested 9 Oct 2026
Claude Haiku 5.5 High effort
How this model behaved
This model worked fast and stopped early, about a minute per attempt. In about half the labs it proved the first rung or two, then ended without completing the exploit or capturing proof. In some of its reports it says the next steps weren't tested, and in a few it declined in its own words. Anthropic's safeguards didn't refuse or block anything. Only proven rungs count, so it has no full solves.
Where it places
Each line is one metric across all 25 configurations. Right is always better; this model is the large dot.
By vulnerability class
How Claude Haiku 5.5 (High) did on each type of vulnerability. Shown as counts, because some classes have only a few labs.
- Authentication / authorisation bypass0 of 28 solved22 of 112 rungs
- Information disclosure0 of 12 solved15 of 48 rungs
- IDOR0 of 11 solved4 of 44 rungs
- SSRF0 of 10 solved0 of 40 rungs
- Business logic0 of 10 solved12 of 40 rungs
- Key or secret exposure0 of 8 solved13 of 32 rungs
- XSS0 of 6 solved3 of 24 rungs
- Token theft0 of 5 solved0 of 20 rungs
Fewer than 5 labs each: one lab changes these a lot, so read them with care.
- Supply chain0 of 4 solved2 of 16 rungs
- Remote code execution0 of 3 solved1 of 12 rungs
- SQL injection0 of 2 solved0 of 8 rungs
- File inclusion0 of 1 solved0 of 4 rungs
Provider performance
Measured per single agent against the provider endpoint used for the run.
- Response time
- 3.44 s median · 10.26 s p95
- Output speed
- 198.6 tok/s
- API success rate
- 100.00%
- Input context
- 8.5K median · 16.7K p95
About this result
Every model runs in the same neutral harness, so results compare the models themselves.
- Harness
- BB Arena neutral harness
- Safety refusals
- None